HiSkill: Hierarchical Skill Graphs¶
Core mechanism evidence. Paper and local evaluation are separate.
Paper information¶
| Field | Value |
|---|---|
| Paper | HiSkill (arXiv 2607.25853) |
| Affiliations | Beijing University of Posts and Telecommunications |
| First public release | 2026-07-28 |
| Authors' code | Author code. |
| Adapter / method | hiskill |
| Local code | src/auto_research/agent_research/ |
Original paper¶
Background and contribution¶
Organize experience into high-level skills, executable AtomicOps, and typed directed edges. Retrieve only a task-relevant subgraph for action grounding.
flowchart LR
G[Skill and AtomicOp graph] --> R[Relevant subgraph] --> A[Ground action]Key figure from the paper¶
Original Figure 1: HiSkill. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\[
G=(V_{\mathrm{skill}}\cup V_{\mathrm{op}},E_{\mathrm{decomp}}\cup E_{\mathrm{temporal}}\cup E_{\mathrm{recovery}}),\quad a_t=\operatorname{Ground}(s_t,G_t).
\]
Paper results¶
The paper beats strong baselines in three interactive environments and lowers inference tokens. No live A/B.
Local reproduction¶
PlanBench mini builds decomposition, transition, and recovery edges and logs reuse. Joint success 1.0000, average cost 0.6900, nodes/edges/AtomicOp reuse 48/60/324. Metrics.
Reproduction limits¶
Graph nodes come from controlled trajectories, not online LLM extraction. Tests diagnose hierarchy and reuse.